The Economist - USA (2022-01-15)

68 Science & technology TheEconomistJanuary15th 2022

stretches 1,000km along the coast of North

America, from Vancouver Island in Canada

to  northern  California.  It  is  the  boundary

between  the  Explorer,  Juan  de  Fuca  and

Gorda  plates  to  the  west,  and  the  North

American  plate  to  the  east.  Steady  move­

ment  of  the  latter  plate  over  the  former

three  generates  a  slow­slip  event  every  14

months  or  so,  and  geophysicists  have  re­

corded  this  activity  in  detail  since  the

1990s. That means there are plenty of com­

plete  cycles  of  data—and  the  machine­

learning  system  trained  on  these  by  Dr

Johnson  was  able  to  “hindcast”  past  slow

slips  based  on  the  seismic  signals  which

preceded  them,  “predicting”  when  they

would  happen  to  within  a  week  or  so  of

when they had occurred in reality.

The next test of the technique, yet to be

executed,  will  be  an  actual  forecast  of  a

slow­slip event. But even without this hav­

ing happened, Dr Johnson’s slow­slip pro­

ject  suggests  that  machine­learning  tech­

niques  do  indeed  work  with  seismic

events, and might thus be extended to in­

clude earthquakes if only there were a way

to compensate for the lack of data. To pro­

vide  such  compensation,  he  and  his  col­

leagues are applying a process called trans­

fer  learning.  This  operates  with  a  mixture

of simulated and real­world information.

Getting real

“Lab  quakes”  are  miniature  earthquakes

generated on a laboratory bench by squeez­

ing  glass  beads  slowly  in  a  press,  until

something  suddenly  gives.  This  has

proved  a  useful  surrogate  for  stick­slip

movement.  Dr  Johnson’s  team  have  creat­

ed  a  numerical  simulation  (a  computer

model that captures the essential elements

of  a  physical  system)  of  a  lab  quake  and

trained their machine­learning system on

it, to see if it can learn to predict the course

of the surrogate quakes. 

The result is moderately successful. But

what really makes a difference is boosting

the trained system with extra data from ac­

tual experiments—in other words, transfer

learning.  The  combination  of  simulated

data  fine­tuned  with  a  pinch  of  the  real

thing  is  markedly  more  effective  at  pre­

dicting when a lab quake will occur.

The next step towards earthquake fore­

casting will be to apply the same approach

to a real geological fault, in this case prob­

ably the San Andreas. A machine­learning

system  will  be  trained  on  data  from  a  nu­

merical  simulation  of  the  fault,  plus  the

half­cycle’s worth of live data available. Dr

Johnson’s team will see if this is enough to

hindcast events not included in the train­

ing  data.  He  mentions  the  magnitude­six

Parkfield  earthquake  in  2004—a  slippage

of the San Andreas that did minimal dam­

age,  but  was  extremely  well  studied—as

one possible target. 

At present Dr Johnson’s aspirations are

limited to predicting the timing of an im­

minent quake. A full prediction would also

need  to  include  whereabouts  along  the

fault it was going to happen and its magni­

tude. However, if timing can indeed be pre­

dicted, that will surely stimulate efforts to

forecast these other criteria, as well. 

He  hopes  for  initial  results  in  the  next

three  to  six  months,  but  cautions  that  it

might take longer than that. If those results

are  indeed  promising,  though,  there  will

no doubt be a rush of other teams around

the world attempting to do likewise, using

historical data from other earthquake­pro­

ducing faults in order to validate the tech­

nique.  That,  in  turn,  should  improve  the

underlying model.

If  it  all  comes  to  naught,  nothing  will

have been lost, for Dr Johnson’s work will

certainlyprovidea betterunderstandingof

thephysicsofbigearthquakes,andthatis

valuableinandofitself.But,ifitdoesnot

cometonaught,andinsteadcreatessoft­

ware capable of predicting when big

quakeswillhappen,thatreallywouldbean

earth­shakingdiscovery.n

Climatechange

Unfrozen North

A

quarter of the  northern  hemi­

sphere’s land is covered by permafrost,

defined as ground that remains at or below

0°C  for  at  least  two  years  in  succession.

Most  of  this  is  above  the  Arctic  Circle,  a

part of the world that is warming at a rate

double the global average, with significant

consequences  for  the  rest  of  the  planet.

Arctic  permafrost  is  thought  to  contain

some 1.7trn tonnes of carbon, most of it in

frozen  organic  matter.  That  is  double  the

amount  of  the  stuff  currently  residing  in

the  atmosphere.  Rising  temperatures

mean that much of this material may turn

into  carbon  dioxide  and  methane  as  the

ground thaws and micro­organisms get to

work.  That  will  drive  further  warming,

causing  a  feedback  loop  of  more  melting

and yet more greenhouse­gas emission.

These  risks  are  re­emphasised  in  a  pa­

per just published in Nature Reviews Earth

and Environment. It warns that warming of

the  top  three  metres  of  permafrost  alone

could result in the release of 624m tonnes

of carbon a year by 2100, a figure similar to

the  current  emissions  of  Canada  or  Saudi

Arabia.  But  a  thawing  Arctic  poses  other,

more immediate, problems. Another paper

published  in  the  same  journal  highlights

the threat posed to circumpolar infrastruc­

ture as the ground beneath it thaws.

Thawing  permafrost  is  a  particularly

unpredictable  environment  on  which  to

build.  As  its  ice  content  changes  and  the

volume  of  liquid  water  increases,  the  soil

can  experience  vertical  movements  of  up

to  40cm  a  year  and  its  capacity  to  bear

weight drops dramatically. This can lead to

landslides, to the subsidence of individual

buildings, and to the appearance of cracks

and  deformities  in  long,  linear  structures

such as roads and pipelines.

The  conclusions  drawn  by  lead  author

Jan Hjort, of the University of Oulu, in Fin­

land,  are  stark.  Of  the  120,000  buildings,

40,000km  of  roads  and  9,500km  of  pipe­

lines  currently  built  on  permafrost,  up  to

half are expected to be at high risk by 2060.

By then, he estimates, the bill for mainte­

nance could exceed $35bn dollars a year.

Russia  is  the  country  most  threatened

Arctic infrastructure is threatened by

rising temperatures

Nothing is permanent

Permafrostcoverage, 2050 forecast*

Expectedtoexist Thawedpermafrost

CANADA RUSSIA

Arctic

Circle

Source:NatureReviewsEarthandEnvironment

*UnderRepresentativeConcentrationPathway (RCP) 4.

†Roads,railways,pipelines,ports,airportsand buildings

North Pole

Glacier

CANADA RUSSIA

Alaska

(US)

Alaska

(US)

Arctic

Circle

North Pole

Glacier

Permafrost degradation risk to infrastructure†

2050 forecast* High Medium Low

The Richard Casement internship. We invite

applications for the 2022 Richard Casement

internship. We are looking for a would-be journalist

to spend three months of the summer working on

the newspaper in London (covid-19 permitting;

otherwise remotely), writing about science and

technology. Applicants should compose a letter

introducing themselves and an article of about 600

words that they think would be suitable for

publication in the Science & technology section. The

successful candidate will receive a stipend of £2,000

a month. Applications must reach us by midnight on

January 28th. They should be sent to:

[email protected]